Predicting Seabed Mud Content across the Australian Margin: Performance of Machine Learning Methods and their combinations with Ordinary Kriging and Inverse Distance Squared

In 2008, the performance of 14 statistical and mathematical methods for spatial interpolation was compared using samples of seabed mud content across the Australian Exclusive Economic Zone (AEEZ), which indicated that machine learning methods are generally among the most accurate methods. In this study, we further test the performance of machine learning methods in combination with ordinary kriging (OK) and inverse distance squared (IDS). We aim to identify the most accurate methods for spatial interpolation of seabed mud content in three regions (i.e., N, NE and SW) in AEEZ using samples extracted from Geoscience Australia's Marine Samples Database (MARS). The performance of 18 methods (machine learning methods and their combinations with OK or IDS) is compared using a simulation experiment. The prediction accuracy changes with the methods, inclusion and exclusion of slope, search window size, model averaging and the study region. The combination of RF and OK (RFOK) and the combination of RF and IDS (RFIDS) are, on average, more accurate than the other methods based on the prediction accuracy and visual examination of prediction maps in all three regions when slope is included and when their searching widow size is 12 and 7, respectively. Averaging the predictions of these two most accurate methods could be an alternative for spatial interpolation. The methods identified in this study reduce the prediction error by up to 19% and their predictions depict the transitional zones between geomorphic features in comparison with the control. This study confirmed the effectiveness of combining machine learning methods with OK or IDS and produced an alternative source of methods for spatial interpolation. Procedures employed in this study for selecting the most accurate prediction methods provide guidance for future studies.

2008年，一项研究以澳大利亚专属经济区（Australian Exclusive Economic Zone, AEEZ）内的海底泥质含量样本为数据基础，对比了14种统计与数学空间插值方法的性能，结果表明机器学习方法总体位列精度最优的方法之列。本研究进一步测试了机器学习方法分别与普通克里金（ordinary kriging, OK）、反距离平方法（inverse distance squared, IDS）结合后的插值性能。我们旨在通过从澳大利亚地球科学局（Geoscience Australia）海洋样本数据库（Marine Samples Database, MARS）中提取的样本，识别澳大利亚专属经济区内三个区域（即北部、东北部和西南部）海底泥质含量空间插值的最优方法。本研究通过模拟实验对比了18种方法（含机器学习方法及其与OK或IDS的组合方案）的性能。预测精度会随插值方法选择、坡度因子的纳入与否、搜索窗口大小、模型平均策略以及研究区域的不同而发生变化。当纳入坡度因子、且两类组合模型的搜索窗口分别设置为12和7时，随机森林（Random Forest, RF）与普通克里金的组合（RFOK）、随机森林与反距离平方法的组合（RFIDS）在三个研究区域内，基于预测精度及预测图目视评估结果，均展现出优于其余方法的平均精度。对这两种精度最优的方法的预测结果进行平均集成，可作为空间插值的备选方案。本研究确定的方法可将预测误差最高降低19%，且相较于对照方案，其预测结果能够清晰刻画地貌特征间的过渡带。本研究验证了机器学习方法与OK或IDS结合的有效性，并为空间插值任务提供了可选的方法体系。本研究中用于筛选最优预测方法的流程，可为后续相关研究提供参考借鉴。